Hi,
This is the my report for week 3 of GSoC.
My goal for the previous week was to implement a basic dependency
tracking framework used to construct dependency DAG in codegen. And I
have finished the task. I also wrote a very simple scheduling policy,
in which instructions with their dependencies met are put into a FIFO,
waiting to schedule. This scheduling policy is definitely not a good
one, without taking latency into account, not to mention register
pressure. However, some simple testing yields interesting result.
Frame time of pixmark piano changed from 127ms to 121ms with
scheduling on. Also, I saw improvements in IPC in some programs that I
wrote. I will put my code on github tomorrow.
Now that dependency tracking is nearly done, my focus for next several
weeks will be implementing some scheduling policies. For the next
week, I plan first to consider latency. This is further divided into
two problems: latency model and how to take advantage of latency
information. I'll see if I can make use of current code for latency
calculation, and try different ways such as top-down or bottom up, and
evaluate their effectiveness. If I can finish them this week, I'll
investigate register pressure.
Regards,
Boyan Ding
